How AI Search Engines Choose Sources

Modern AI search systems use an architecture called Retrieval-Augmented Generation (RAG). When a user submits a query, the system performs a two-stage process: first, it retrieves relevant passages from an indexed content corpus; second, it uses a language model to synthesise those passages into a coherent response.

The retrieval stage is where source selection occurs. The system embeds the query as a vector — a mathematical representation of its meaning — and compares that vector against embeddings of indexed passages to identify the most semantically similar content. This semantic similarity search is more sophisticated than keyword matching: two passages can be retrieved as relevant even if they share no keywords with the query, as long as their semantic content addresses the same intent.

After initial semantic retrieval, a re-ranking stage applies additional quality signals to filter and prioritise the retrieved passages. It is at this re-ranking stage that signals like factual specificity, source credibility, passage clarity, and structural quality determine which passages are ultimately included in the generated response.

For the full GEO context, return to the complete GEO guide.

The most important concept in understanding AI source selection is that AI systems retrieve passages, not pages. A single page on your website might contain 20 passages. The AI system evaluates each passage independently and selects the passage most relevant to the query — regardless of what the rest of the page says.

This has two significant implications. First, a page with one excellent passage and nineteen weak ones can still earn AI citations for the topic addressed by that excellent passage. Second, a page that is topically comprehensive but lacks any self-contained, extractable passages may rank highly in traditional search while earning zero AI citations.

This is precisely why GEO requires a different content architecture from pure SEO. The goal is not just to build comprehensive pages — it is to build pages where every paragraph is independently valuable and extractable.

This passage-level retrieval mechanism is also the basis of Google's Passage Ranking system, which indexes individual passages for traditional search. Read How Passage Ranking Affects AI Search Visibility for a detailed analysis of how passage ranking and AI retrieval interact.

While all major AI search systems use RAG-based retrieval with similar core signals, they differ in their source corpora, re-ranking priorities, and citation display behaviour.

Google AI Overviewsdraw from Google's indexed web corpus with direct influence from Google's Knowledge Graph and quality systems (E-E-A-T, Helpful Content). Pages that already rank well in Google Search are more likely to appear in AI Overviews, but ranking is not sufficient — passage extractability remains the deciding factor for inclusion.

Perplexity uses a real-time web search layer combined with RAG. Its source selection is more transparent — it cites sources inline and gives users direct access to attribution. Perplexity prioritises recent content and tends to favour pages with clear authorship, specific data points, and well-structured information over long-form content without clear passage demarcation.

ChatGPT Search(via Bing integration) applies Microsoft's index with OpenAI's language model layer. It tends to favour established domains and authoritative sources on specific topics. Its citation behaviour is less consistent than Perplexity, but it responds to the same core GEO signals: specificity, structure, and source credibility.

A unified GEO strategy that applies the six core signals — semantic relevance, factual specificity, extractability, source credibility, freshness, and structural quality — performs across all three platforms simultaneously.